A Graph Neural Network with Type-Feature Attention for Node Classification on Heterogeneous Graphs

Abstract

Heterogeneous graphs are emerging as a prevalent form of data representation to capture complex structures and different relationships between a set of different types of objects in diverse disciplines. Node classification on heterogeneous graphs is a basic and critical task that remains unaddressed until the present day. Graph Neural Network is a powerful tool and has demonstrated remarkable performance in various tasks on graphs. However, most existing graph neural networks are based on the homophily assumption, which may be unsuitable for heterogeneous graphs. In this paper, we propose a graph neural network with type-feature attention mechanism to solve the problem of node classification on heterogeneous graphs. As a heterogeneous graph is composed of a group of edges between different types of nodes, it is reasonable to assume that each type of edge plays a different role in message propagation with different importance. An attention mechanism that considers the edge type and the features of the end nodes of the corresponding edge is built and incorporated into the process of message propagation of the graph neural network, by which the different heterogeneous information of nodes and edges is used jointly in solving the problem of node classification. We evaluate the proposed method on two public real-world heterogeneous graphs and the experimental results demonstrate the effectiveness of the proposed method.